app.py

import gradio as gr
import os
import json
from datetime import datetime
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
import torch
import soundfile as sf

MODEL_NAME = "mistralai/Voxtral-Small-24B-2507"

processor = AutoProcessor.from_pretrained(MODEL_NAME)
model = AutoModelForSpeechSeq2Seq.from_pretrained(MODEL_NAME)

def transcribe(audio_file, metadata_file):
    # Lade Audio
    audio_input, sample_rate = sf.read(audio_file)
    inputs = processor(audio_input, sampling_rate=sample_rate, return_tensors="pt")
    with torch.no_grad():
        generated_ids = model.generate(**inputs)
    transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

    # Lade Metadaten
    metadata = None
    if metadata_file is not None:
        metadata = json.load(metadata_file)

    # Ergebnis
    result = {
        "transcription": transcription,
        "model": MODEL_NAME,
        "timestamp": datetime.now().isoformat(),
        "metadata": metadata
    }
    return json.dumps(result, ensure_ascii=False, indent=2)

demo = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(type="filepath", label="Audio/Video-Datei (mp3, wav, mp4, etc.)"),
        gr.File(label="Passende Metadata JSON-Datei")
    ],
    outputs=gr.Textbox(label="Transkriptions-JSON"),
    title="Voxtral Transkription mit Metadaten"
)

if __name__ == "__main__":
    demo.launch()